Multi-scale residual aggregation feature network based on multi-time division for motion behavior recognition

The existing behavior recognition models based on the deep convolutional neural network have some problems, such as feature extraction with a single scale and insufficient feature utilization in the middle level. In this paper, we propose a multi-scale residual aggregation feature network based on multi-time division for behavior recognition. Through the sampling form of multi-time division, the diversity of behavior depth features is enriched. Firstly, a hybrid extended convolution residual block (HERB) is designed using extended convolution and residual join with different extension coefficients to extract feature information at multiple scales effectively. Secondly, a feature aggregation mechanism (AM) is introduced to solve the problem of insufficient feature utilization in the middle layer of the network. We construct a deep aggregation model that can learn the distribution of complex behavior features to solve the problem of human behavior classification over a long time span. Experiments on behavioral datasets UCF101 and HMDB51 verify the effectiveness of the new algorithm.